论文总字数：24683字

摘 要

大气污染是环境问题的重要部分，精准测量污染气体的浓度对环境保护和环境治理有着重要的意义。

首先本文以朗伯-比尔定律为理论基础，分析了气体吸收的数学模型，主要考虑了背景暗噪声和多种气体共存的情况；介绍了差分吸收光学，通过差分算法，可以基本消除散射的影响，在此基础上，推导分析了最小二乘法在浓度反演中的运用。

接着本文介绍了二氧化硫、一氧化氮、二氧化氮和氨的传统浓度反演。研究了多项式拟合的差分算法；对最小二乘法提出了改进，分析了FFT预处理与特征波段处理对误差的影响；对未知气体的吸收截面估计进行了研究。通过仿真实验，FFT预处理及特征波段处理有利于提高浓度反演性能；但对二氧化氮等吸收特性不明显的气体浓度反演和未知气体谱的估计，表现较差。

随后本文介绍了深度前馈网络在浓度反演中的运用。分析了网络结构，探讨了网络的训练方法和超参数的设定，分析了几种梯度下降法和自适应学习率算法的优劣，对过拟合及Dropout进行了研究。仿真实验表明，采用深度学习的气体浓度反演算法误差性能优于传统多项式拟合差分的最小二乘法，且在未知气体的吸收截面估计方面也有着较好的性能增益。

最后本文对传统浓度反演和深度前馈网络浓度反演进行了性能对比总结，对实验中的不足进行了分析，并对气体浓度分析方法进行了展望。

关键词：朗伯-比尔定律，最小二乘法，深度前馈网络

Abstract

Air pollution is an important part of environmental problems. Accurate measurement of the concentration of polluted gases is of great significance to environmental protection and environmental governance.

Firstly, based on Lambert-Beer's law, the mathematical model of gas absorption is analyzed. The background dark noise and the coexistence of various gases are mainly considered. Differential absorption optics is introduced. The scattering effect can be basically eliminated by differential algorithm. On this basis, the application of least square method in concentration inversion is deduced and analyzed.

Next, the traditional concentration inversion of SO₂, NO, NO₂ and NH₃ is introduced. The difference algorithm of polynomial fitting is studied, the least square method is improved, the influence of FFT pretreatment and characteristic band processing on the error is analyzed, and the estimation of absorption cross section of unknown gas is studied. Through simulation experiments, FFT pretreatment and characteristic band processing are beneficial to improve the performance of concentration inversion, but the performance of gas concentration inversion and estimation of unknown gas spectrum which have not obvious absorption characteristics such as nitrogen dioxide is poor.

Then, the application of depth feedforward network in concentration inversion is introduced. The structure of the network is analyzed, the training methods and the setting of super parameters are discussed, the advantages and disadvantages of several gradient descent methods and adaptive learning rate algorithms are analyzed, and the over-fitting and Dropout are studied. The simulation results show that the error performance of the gas concentration inversion algorithm based on deep learning is better than that of the least square method based on traditional frequency domain difference, and it also has better performance gain in estimating the absorption cross section of unknown gas.

Finally, the performance of traditional concentration inversion and depth feed-forward network concentration inversion is compared and summarized. The shortcomings of the experiment are analyzed, and the gas concentration analysis method is prospected.

KEY WORDS: Beer-Lambert law, least square method, depth feedforward network

目 录

摘 要 I

Abstract II

第一章 绪论 1

1.1 课题背景及研究意义 1

1.2 研究现状 1

1.2.1 气体检测技术的研究现状 1

1.2.2 差分吸收光谱技术的研究现状 2

1.3 本课题研究的主要内容 3

第二章 气体浓度反演原理与模型 4

2.1 朗伯比尔定律 4

2.2 气体吸收的数学模型 4

2.3 差分吸收光学 5

2.4 浓度反演算法 6

第三章 传统浓度反演算法 8

3.1 引言 8

3.2 混合气体反演波段的选择 8

3.3 差分处理方式 10

3.3.1 多项式拟合 10

3.3.2 超参数的选取 10

3.3.3 实验结果 10

3.4 浓度反演算法的改进 11

3.4.1 光谱预处理 11

3.4.2 特征波段处理 14

3.5 未知气体吸收截面的估计 15

第四章 深度学习检测网络 17

4.1 引言 17

4.2 深度学习检测网络 17

4.2.1 前向传播 17

4.2.2 损失函数 19

4.2.3 基于梯度的优化方法 20

4.2.4 参数初始值 21

4.2.5 超参数 22

4.2.6 正则化与Dropout 25

4.3 仿真结果与说明 27

第五章 总结与展望 30

5.1 论文工作总结 30

5.2 展望 30

参考文献 32

致 谢 34

绪论

课题背景及研究意义

随着国家社会经济的发展，环境污染问题逐渐引起了社会的关注。恶劣的大气环境对人们的身体健康造成了严重的危害，诱发各种呼吸道疾病，大气污染治理迫在眉睫。

工业化进程的加速，相关环保技术、法规和意识的缺乏，工业废气排放日益严重，尤其燃煤工业，是造成大气污染的主要原因之一。城市化进程的加速，对便捷出行的需求逐步提升，汽车保有量逐年增加，汽车尾气对环境的影响不可忽略。我国城市大气中，相关污染物指数逐年上升，尤其是SO₂、NO_x等污染气体浓度。

剩余内容已隐藏，请支付后下载全文，论文总字数：24683字